This secondary analysis of a prospective cohort study is reported according to the STROBE guidelines [21]. The institutional review boards at the Universities of Alberta (Pro00117530) and Calgary (pSite-22-0007 Pro00117530) approved this study. Written informed consent was obtained. The data are publicly available (https://wwwn.cdc.gov/nchs/nhanes/Default.aspx and https://www.cdc.gov/nchs/data-linkage/mortality-public.htm).

We used eight cycles of the publicly available National Health and Nutrition Examination Surveys (NHANES; 1999–2000 to 2013–2014) and linked them with the publicly available National Center for Health Statistics (NCHS) 2015 mortality file [22] to create an initial larger cohort (i.e., the BMI cohort). Only six NHANES cycles from 1999–2000 to 2009–2010 had data on c-reactive protein (CRP) and fasting insulin and we used these data to form the primary cohort. NHANES is a series of surveys conducted on non-institutionalized US adults identified through stratified four-stage probability random sampling. Through interviews and physical examinations, the NHANES collects information on demographics, physiological measurements, medical history, medication use, and laboratory tests.

The database was used to assemble two cohorts of adults (age ≥20 years) residing in the United States who had a measure of BMI and vital status ascertained as of December 31, 2015. This first cohort (the BMI cohort) included participants with complete data on biological sex, BMI, and vital status. The second cohort, a subset of the BMI cohort (the primary cohort), included participants with complete data on age, biological sex, BMI, fasting insulin, CRP, and vital status.

The exposures of interest were CRP, fasting insulin. and BMI. If a participant fasted for <8 or >24 h, their fasting insulin and glucose values were not used. CRP was assayed using latex-enhanced nephelometry. Its lower detection limit was 0.2 mg/L. The NHANES quality control and quality assurance protocols met the 1988 Clinical Laboratory Improvement Act requirements.

All-cause mortality was the study outcome. The NCHS used 13 variables to probabilistically ascertain vital status of survey participants: social security number, first name, middle initial, last name, father’s surname, month of birth, day of birth, year of birth, state of birth, state of residence, sex, race, and marital status.

Age, biological sex, cigarette smoking, and ten self-reported chronic conditions (angina, arthritis, diabetes, cancer, chronic heart failure, chronic liver disease, chronic lung disease, coronary artery disease, stroke, and thyroid problems) were included as covariates. Specifically, the participants were asked if a doctor or another health professional ever indicated that they had a particular condition. The ten chronic conditions were selected based on their availability in the datasets, and their associations with obesity. Fasting glucose, and glycated hemoglobin [HbA1c]) were considered in sensitivity analyses along with insulin resistance. Insulin resistance was measured using the Homeostatic Model Assessment [23] (HOMA-IR).

We did the analyses with Stata MP 17.0 (www.stata.com) and reported baseline descriptive statistics as percentages, or medians and inter-quartile ranges, as appropriate. The NHANES uses a four-stage probability random sampling procedure (counties, city blocks, households, then individuals) and they oversample low-frequency demographic subgroups. In order to produce nationally representative results and to account for the sampling design (unequal probability of selection), we applied sampling weights. The variance was estimated using Taylor linearization.

In order to characterize the association between BMI and mortality, we used the larger BMI cohort. Time to mortality was regressed on BMI, parametrized with restricted cubic splines, and interacted with biological sex using Cox regression models. These unadjusted hazard ratios with 95% confidence intervals were plotted against BMI, and underlaid with the distributions of BMI by biological sex. Standard errors by biological sex were calculated.

Using the primary cohort, medians and inter-quartile ranges of fasting insulin and CRP are presented by BMI categories (<18.5, 18.5–<25, 25–<35, 35–<45, ≥45 kg/m2) and by biological sex. Differences by sex were tested with gamma or linear regression, as appropriate. Correlations between fasting insulin, CRP, and BMI were calculated. The distributions of BMI were plotted in pie charts for participants in the top 25th percentiles of fasting insulin and CRP, respectively.

Using the primary cohort, time to mortality was regressed on all three exposures of interest: BMI, fasting insulin, and CRP. A number of parametrizations of these three exposures were explored. The models with the best fit used linear and quadratic terms for BMI, a linear term for fasting insulin, and the natural logarithm of CRP. Model 1 adjusted for BMI, age, biological sex, and cigarette smoking (categorized as yes, no, and missing). Model 2 was further adjusted for fasting insulin and CRP. Model 3 further adjusted for the ten chronic conditions listed earlier.

We did four sensitivity analyses based on the fully adjusted model 3. First, we added glycated hemoglobin to model 3. Second, we added fasting glucose to model 3. Third, we replaced fasting insulin in model 3 with the HOMA-IR. Fourth, we replaced BMI in model 3 with percentage fat mass in further models.

To separate the effects of fasting insulin and CRP on the association of BMI with mortality, we used model 3 to evaluate every permutation of these three exposures, alone or in combination.

To explore whether the extent of inflammation might further modify the associations between fasting insulin, BMI, and mortality, we categorized participants as having CRP ≤ 10 mg/L vs >10 mg/L (none/low vs moderate/high grade of inflammation [24]). We then interacted the presence or absence of a moderate/high grade of inflammation with BMI and fasting insulin (model 4). The natural logarithm of CRP remained in this model as a main effect only.

Finally, we interacted with biological sex with the three exposures in the most adjusted third model (model 5).

All models, using the primary cohort, expressed hazard ratios with 95% confidence intervals for the 1st, 50th, 75th, 95th, and 99th percentiles of the three exposures as compared to a referent category of the 5th percentile. The 5th percentile was selected as the referent as a BMI of 20 kg/m2 corresponds with a so-called “healthy” BMI. We determined that the proportional hazard assumption was satisfied by examining plots of the log-negative-log of within-group survivorship probabilities versus log-time. The threshold two-sided p for statistical significance was set at 0.05.